Learn how a real AI team works.

The Full AI Stack, Explained (October 2026)

Grok Bot University · Piece 00 · by Fabian de la Fuente

When someone tells me "I use AI," they almost always mean "I use ChatGPT." Which is great! But it's a bit like saying "I own a car" because you're holding a steering wheel. Under that wheel there's an engine, a transmission, fuel, brakes… and somebody making sure nothing catches fire.

AI in 2026 works the same way. It isn't "a chatbot." It's a stack of 10 layers, one on top of the other. This infographic puts them all on one page, with 83 tools, so you can see the whole car and not just the wheel.

You don't need all 83. I don't use all 83. Nobody does (and if someone says they do, be suspicious). The point is that the next time you hear a weird tool name, you'll know which floor of the building it lives on.

How to read the image

• Left: 10 colored bands. Each one is a layer.
• Right: a white box with that layer's tools (logo + name + a two- or three-word job description).
• Red "NEW!" stamps: what changed the most this year (voice and phone agents, agents that use a computer, automated routines).
• Red dashed borders: Grok Bot, Grok and Grok Bot Routines. They're highlighted because that's what I run my agent team on (and Grok Bot made the infographic).
• An honesty note: the image says "GPT-5" next to OpenAI. As of today (Oct 14, 2026), OpenAI's current lineup is the GPT-6 family. AI moves so fast that infographics expire like milk.

About prices: I only list prices I checked today on each company's official page (US dollars, before tax). If I couldn't confirm a price, I left it out. A blank beats a made-up number.

Layer 1 · Assistants + personal agents 🟨

In plain words: the part you actually see. You type or talk, it helps. The newest ones don't just answer: they get tasks done.

• Grok Bot: your own team of agents, each with a role. docs.x.ai/x-subscription/grok-bot · Included in The X Subscription: Premium $30/month, Super $100, Ultra $200 (the $8 Plus plan doesn't include it). Source: x.ai/pricing
• ChatGPT: OpenAI's assistant, which now runs tasks for you. chatgpt.com/pricing · Plus $20/month; Pro from $100/month.
• Claude: Anthropic's assistant, great with writing and documents. claude.com/pricing · Pro $20/month ($17 billed annually); Max from $100/month.
• Gemini: Google's assistant, built into Gmail, Docs and friends. gemini.google/subscriptions · Free tier available; Google AI Ultra is $99.99/month (or $199.99 for more usage). I couldn't confirm the Google AI Pro price today.
• Manus: an autonomous agent. Give it a goal and it goes after it. manus.im
• Perplexity: a search engine that answers you, with sources. perplexity.ai
• Microsoft Copilot: the AI inside Office and Windows. copilot.microsoft.com
• Meta AI: the AI inside WhatsApp and Instagram. meta.ai

Layer 2 · Language models (LLMs) 🟪

In plain words: the brain. Layer 1 is the face; these models do the thinking behind it. Companies charge per million tokens (a token is a chunk of a word).

• Grok (xAI): xAI's model family, the brain behind the Grok assistant. docs.x.ai/docs/models · Grok 4.7: $2 per million input tokens, $6 per million output.
• OpenAI: the GPT models doing the thinking behind ChatGPT; any app can use them through the API. platform.openai.com/docs/pricing · GPT-6.1 Sol: $2 in / $10 out per million.
• Claude (Anthropic): the models behind the Claude assistant, popular for writing and coding. claude.com/pricing · Sonnet 5.5: $2 in / $10 out per million.
• Gemini (Google): Google's models, which can make sense of text, images, audio and video. deepmind.google/models/gemini
• DeepSeek: Chinese models that surprised everyone by thinking so well for so little money. deepseek.com
• Qwen (Alibaba): a huge family of models, from tiny to giant, many of them open. qwen.ai
• Mistral: Europe's big bet, with efficient models, several of them open. mistral.ai
• Llama (Meta): Meta's open models, the foundation for a ton of projects. llama.com
• Kimi (Moonshot): a model built for long documents and agent tasks. kimi.com
• Several of these last ones come in open versions you can download and run yourself. I didn't check their API prices today.

Layer 3 · Inference + access 🟧

In plain words: if the model is the engine, this is the gas station. These services run models for you (faster or cheaper), or let you run them on your own machine.

• Hugging Face: the giant library of models, where people share and download models and datasets, mostly for free. huggingface.co
• OpenRouter: one key, many models. openrouter.ai
• Together AI: a cloud full of models. You rent open models through an API without buying your own hardware. together.ai
• Fireworks AI: fast inference. It runs open models in its cloud so your app gets answers in the blink of an eye. fireworks.ai
• Groq (with a q, not to be confused with Grok): custom chips (LPUs) for very fast answers. groq.com
• Cerebras: chips the size of a whole silicon wafer. They run models ridiculously fast, and you can use them through Cerebras's cloud. cerebras.ai
• Ollama: run models on your own computer. ollama.com
• vLLM: the engine a lot of people use to serve models: open-source software that runs a model on your servers and handles lots of users at once without choking. docs.vllm.ai
• Prices depend on the model, and I didn't check them today.

Layer 4 · Voice + phone agents 🩷 NEW!

In plain words: AI that listens and talks, over the phone or online. This is where Martha S. lives, my assistant who now makes and answers calls (Piece A02 has the full walkthrough).

• Twilio: the bridge to the phone network; ConversationRelay gives your AI ears and a voice. twilio.com/docs/voice/conversationrelay · ConversationRelay $0.07/min; US/Canada outbound calls $0.014/min; inbound to a local number $0.0085/min; local number $1.15/month. Source: twilio.com/en-us/voice/pricing/us
• Deepgram: speech to text. deepgram.com/pricing · Nova-3 streaming $0.0048/min (promotional rate); $200 credit for new accounts.
• ElevenLabs: voices that sound human. elevenlabs.io/pricing/api · Flash/Turbo via API $0.04 per 1,000 characters; plans run from Free to Starter at $6/month and Creator at $22/month.
• OpenAI Realtime: one model that listens and speaks directly, voice to voice. platform.openai.com/docs/pricing · gpt-realtime audio: $32 per million input tokens, $64 per million output.
• Vapi: a platform for building voice agents. vapi.ai/pricing · $0.05/min hosting plus each provider's cost.
• LiveKit Agents: open-source framework for voice and video agents. livekit.io/pricing · Their cloud: free plan; Ship from $50/month.
• Cartesia: very fast text to speech. cartesia.ai/pricing · Free plan; Pro $5/month; managed agents at $0.06/min.
• Retell AI: a ready-made AI call center. retellai.com/pricing · $0.07 to $0.31/min.

Layer 5 · Computer use + browser 🩵 NEW!

In plain words: agents that move the mouse, click and fill in forms the way you would. Handy when a website has no official "plug" for software.

• Computer Use (Claude): Claude driving a computer. anthropic.com/news/3-5-models-and-comput…
• ChatGPT Atlas: OpenAI's browser. chatgpt.com/atlas
• Comet: Perplexity's browser. perplexity.ai/comet
• Browserbase: cloud browsers for your agents. browserbase.com
• Playwright: automate browsers with code. playwright.dev
• Browser Use: open source, lets an AI drive a browser. browser-use.com
• Stagehand: an SDK for browser agents. stagehand.dev
• E2B: safe sandboxes where AI can run code. e2b.dev
• Prices: not checked today.

Layer 6 · Automation + routines 💚 NEW!

In plain words: the stuff that keeps working while you sleep. "Every Monday at 8, do this." "When an email like that arrives, do that."

• n8n: visual AI workflows. Drag boxes around, wire them together, and you've got an automation between your apps. n8n.io/pricing · Starter €20/month and Pro €50/month (billed annually). There's also a version you can host yourself.
• Zapier: connects thousands of apps, now with agents. zapier.com/pricing · Free with 100 tasks/month; Pro from $19.99/month billed annually ($29.99 monthly).
• Make: visual scenarios. You draw how your apps pass data around ("when this comes in, do that") and Make runs it. make.com/en/pricing · I couldn't confirm their price today.
• Grok Bot Routines: scheduled tasks for my agents, like the digest I get every morning. I didn't find a separate price on the pricing page; Grok Bot usage draws from the plan's weekly pool.
• Postiz: schedules social posts; we use it for Facebook and Instagram. postiz.com/pricing · Standard $29/month ($23 billed annually), with a 7-day free trial.
• Temporal: crash-proof workflows. If a long process dies halfway through, it picks up right where it left off. temporal.io
• Inngest: background jobs. It runs scheduled or long tasks, and retries them if they fail, so your app doesn't sit there waiting. inngest.com
• Pipedream: integrations with code. Small snippets that connect apps and fire whenever something happens. pipedream.com

Layer 7 · Agent frameworks 🔵

In plain words: LEGO kits for developers who want to build their own agents.

• LangGraph (LangChain): build an agent as a flowchart of steps, with memory and the ability to loop back when something goes wrong. langchain.com/langgraph
• AI SDK (Vercel): for adding AI chat and agents to a web app in TypeScript. ai-sdk.dev
• Agents SDK (OpenAI): OpenAI's kit for agents that use tools and hand tasks off to each other. openai.github.io/openai-agents-python
• Agent SDK (Claude): Anthropic's kit for building your own agents with Claude. docs.claude.com/en/docs/agent-sdk/overview
• ADK (Google): Google's agent kit, which plays especially nicely with Gemini. google.github.io/adk-docs
• CrewAI: multi-agent teams. crewai.com
• LlamaIndex: agents that work with documents. llamaindex.ai
• Mastra: for TypeScript. Agents, workflows and memory in one package. mastra.ai
• Pydantic AI: typed (tidier) agents. In Python, it makes sure the AI answers in exactly the format your code expects. ai.pydantic.dev

Layer 8 · Tools + protocols 🟢

In plain words: the plugs. How an agent connects to your email, a database, or another agent.

• MCP: the standard for connecting tools to an AI (think "USB for agents"). modelcontextprotocol.io
• A2A: lets one agent talk to another. a2a-protocol.org
• Agent Skills: skills packaged in a SKILL.md file: instructions that teach an agent one specific job, loaded only when it needs them. agentskills.io
• Composio: hundreds of ready-made integrations. composio.dev
• Exa: a search engine built for AIs. exa.ai
• Tavily: a search API. Your agent asks a question and gets back clean web results, ready to use. tavily.com
• Firecrawl: turns web pages into text an AI can read. firecrawl.dev
• LlamaParse: reads messy PDFs and documents. llamaindex.ai/llamaparse

Layer 9 · Memory + vector data 🟫

In plain words: how your AI remembers things. A "vector database" stores the meaning of text, so the AI can find "stuff like this" even when you use different words.

• Mem0: a memory layer. It gives your agent long-term memory, so it remembers your preferences from one conversation to the next. mem0.ai
• Zep: graph-shaped memory. It stores facts and how they connect (who, what, when). getzep.com
• Letta: agents that keep their state. They still know where they left off after you close the session. letta.com
• Postgres + pgvector: the classic database, now with vectors. github.com/pgvector/pgvector
• Pinecone: managed vectors. A vector database in the cloud; they run it, you just use it. pinecone.io
• Qdrant: a vector database you can run yourself or use in their cloud. qdrant.tech
• Chroma: a lightweight vector database that's easy to start experimenting with. trychroma.com (both open source)
• S3 Vectors (AWS): stores vectors right inside Amazon's S3 storage; handy when you have a mountain of data. aws.amazon.com/s3/features/vectors
• Weaviate: hybrid search. It mixes searching by meaning with searching by exact words. weaviate.io

Layer 10 · Evaluation + observability 🔷

In plain words: quality control. Is my agent answering well? What's it costing me? Where did it go wrong? The least glamorous layer in the stack, and the one that separates hobbyists from pros.

• Langfuse: open source; logs every step your agent takes: what it was asked, what it answered, and what it cost. langfuse.com
• LangSmith: LangChain's platform for tracing and testing what your agents do. langchain.com/langsmith
• Braintrust: evaluations that compare versions of your AI, so you know whether the new one is actually better. braintrust.dev
• Phoenix (Arize): open source; lets you look inside every answer and grade it. phoenix.arize.com
• Promptfoo: "attack" tests to see whether your AI can be tricked. promptfoo.dev
• Opik (Comet): open source; logs and evaluates what your agents do. comet.com/site/products/opik
• Ragas: metrics for AI search. It checks whether answers stick to the documents the AI actually found. docs.ragas.io
• DeepEval: tests for LLMs. Think automated software tests, but for your AI's answers. deepeval.com

So what do I actually use?

Less than you'd think:

• Layer 1: Grok Bot, home to my 18-agent team (Piece A01).
• Layer 4: Twilio, Deepgram, ElevenLabs and an OpenAI model so Martha S. can talk on the phone, plus OpenAI Realtime as a second mode (Piece A02).
• Layer 6: Grok Bot Routines for the things that run on their own, and Postiz to schedule posts.

The rest is on my radar. The lesson: you don't need the whole stack. You need to know which floor you're missing.

Which layer are you on? 👇 Save this and share it with someone who's just getting started with AI.

Prices checked October 14, 2026 on the official pages cited above; they can change. Where there's no price, I couldn't confirm it today. Brands belong to their owners · No affiliation · Made with Grok Bot 🤖